AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

OrthoFuse: Training-free Riemannian Fusion of Orthogonal Style-Concept Adapters for Diffusion Models

Ali Aliev, Kamil Garifullin, Nikolay Yudin, Vera Soboleva, Alexander Molozhavenko, Ivan Oseledets, Aibek Alanov, Maxim Rakhuba

CVPR 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
merge style and concept adapters with OrthoFuse instead of K-LoRA or ZipLoRASDXL orthogonal adapters · concept-style fusionCLIP style similarity: 0.49 -> 0.61Improved
merge with OrthoFuse instead of K-LoRA or ZipLoRASDXL · concept-style fusionconcept CLIP / DINO similarity: CLIP 0.76 -> 0.68; DINO 0.64 -> 0.51Hurt
merge with OrthoFuse instead of K-LoRA r = 64SDXL · concept-style fusiongeometric mean (style, CLIP): 0.61 -> 0.64Improved
fuse training-free with OrthoFuse instead of joint orthogonal trainingSDXL · concept-style fusionCLIP style similarity: 0.48 -> 0.61Improved
merge training-free instead of joint trainingSDXL · concept-style fusionmerging time: 1.5 hours -> < 1 secImproved
increase LoRA rank of baselines from 8 to 64ZipLoRA · concept-style fusionDINO: 0.55 -> 0.64Improved
increase LoRA rank of baselines from 8 to 64K-LoRA · concept-style fusionCLIP style similarity: 0.46 -> 0.49Improved
add spectra restoration (eigenvalue rotation) after geodesic interpolation (ablation)OrthoFuse · SDXL concept-style fusionstyle expressiveness (eigenvalue spectrum)Improved
set fusion parameter t to 0.6 (ablation)OrthoFuse · SDXL concept-style fusionstyle-concept balance: t = 0.6Improved
set phase multiplier midpoint phi0 = 2 (ablation)OrthoFuse · SDXL concept-style fusionfusion quality: phi0 = 2Improved
compare against K-LoRA in human preference on concept preservationOrthoFuse · user study (65 participants, 1460 comparisons)concept preservation preference: 48% vs 52%No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy